Increasing TCP initial cwnd

Increasing TCP initial cwnd

[Yair Gottdenker]

Hi all,

I am working on research project to control the sender initial
congestion window size. I am trying to allow user space to set the
initial congestion window size but with no luck. The sender always
sends just 4 packets disregarding from the snd_cwnd.

I am working on kernel version 2.6.31.3.

I made the following changes:
1. tcp_ipv4.c -> in function tcp_v4_init_sock, changed from
tp->snd_cwnd = 2 to tp->snd_cwnd = user_space_initial_cwnd.
2. tcp_output.c -> tcp_init_cwnd - to always return user_space_initial_cwnd.
3. tcp_cong.c -> tcp_slow_start- to always return user_space_initial_cwnd.

Any help will be more than welcome.

Yair

Re: Increasing TCP initial cwnd

[Ilpo Järvinen]

On Mon, 26 Oct 2009, Yair Gottdenker wrote:

> I am working on research project to control the sender initial
> congestion window size. I am trying to allow user space to set the
> initial congestion window size but with no luck. The sender always
> sends just 4 packets disregarding from the snd_cwnd.
> 
> I am working on kernel version 2.6.31.3.
> 
> I made the following changes:
> 1. tcp_ipv4.c -> in function tcp_v4_init_sock, changed from
> tp->snd_cwnd = 2 to tp->snd_cwnd = user_space_initial_cwnd.
> 2. tcp_output.c -> tcp_init_cwnd - to always return user_space_initial_cwnd.
> 3. tcp_cong.c -> tcp_slow_start- to always return user_space_initial_cwnd.
> 
> Any help will be more than welcome.

How about trying tcp_init_metrics in tcp_input.c?

-- 
 i.

Re: Increasing TCP initial cwnd

[Ilpo Järvinen]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 1127 bytes --]

On Mon, 26 Oct 2009, Ilpo Järvinen wrote:

> On Mon, 26 Oct 2009, Yair Gottdenker wrote:
> 
> > I am working on research project to control the sender initial
> > congestion window size. I am trying to allow user space to set the
> > initial congestion window size but with no luck. The sender always
> > sends just 4 packets disregarding from the snd_cwnd.
> > 
> > I am working on kernel version 2.6.31.3.
> > 
> > I made the following changes:
> > 1. tcp_ipv4.c -> in function tcp_v4_init_sock, changed from
> > tp->snd_cwnd = 2 to tp->snd_cwnd = user_space_initial_cwnd.
> > 2. tcp_output.c -> tcp_init_cwnd - to always return user_space_initial_cwnd.
> > 3. tcp_cong.c -> tcp_slow_start- to always return user_space_initial_cwnd.
> > 
> > Any help will be more than welcome.
> 
> How about trying tcp_init_metrics in tcp_input.c?

Also, now that I looked. I don't understand what interesting you find in 
tcp_slow_start to modify? That certainly has nothing to do with initial 
window. ....Then you also might want to play with RW but that 
is out-of-scope of your question, just mentioning it for completeleness.

-- 
 i.

Re: Increasing TCP initial cwnd

[Yair Gottdenker]

Thanks,

I already had a look in it. Their changes that are related to changing
the initial cwnd  (snd_cwnd) are in several places in the patch:
1. tcp_ipv4.c in function tcp_init_cwnd. since currently I am working
with no metric this is irrelevant.
2. tcp_ipv4.c: in function tcp_v4_init_sock, which I applied
3. tcp_minisocks.c - which I applied
. They patched kernel 2.6.18.8 which is very old and there were a lot
of changes since that in the kernel tcp layer. I didn't test their
patch but applying the same logic to 2.6.31.3 doesn't seems to change
initial window size.

It is clear that in order to correctly make the change to the initial
cwnd, one should apply changes to many more scenarios like setting the
initial cwnd after idle time and more. I want to proceed step by step,
first the naive changes to make it work.




On Mon, Oct 26, 2009 at 4:38 PM, Ilpo Järvinen
<ilpo.jarvinen@helsinki.fi> wrote:
> On Mon, 26 Oct 2009, Yair Gottdenker wrote:
>
>> Thanks for the response.
>> I disabled metric by: net.ipv4.tcp_no_metrics_save = 1 in sysctl.conf.
>> But just after the call to tcp_init_metrics in tcp_input.c there is
>> also a call to: tcp_init_congestion_control which calls the congestion
>> control algorithm that is in used. I created my own and in the init
>> function I set the snd_cwnd to 10 and still there is no change and
>> only 4 segments are initially sent.
>> You are right tcp_slow_start has nothing to do with the cwnd size, I
>> mistakenly thought that the initial flow enters that function.
>
> Hopefully you don't mess up anything in there.
>
> Please see from http://www.cs.helsinki.fi/research/iwtcp/kernel-patch/
> I haven't had time/interest to bring it up to date but at least there's
> very clearly shown how the initial window can be configured through a
> sysctl (though don't try to patch it to late kernels, it would just be
> a never ending pain for you :-)).
>
>> What do you mean RW? do you mean tcp read/write buffer size?
>
> RW = Restart Window (RFC2681 IIRC, nowadays also in RFC 5681 I think).
> ...It is related to slow-start after idle.
>
> --
>  i.
>

Re: Increasing TCP initial cwnd

[Ilpo Järvinen]

On Mon, 26 Oct 2009, Yair Gottdenker wrote:

> I already had a look in it. Their changes that are related to changing
> the initial cwnd  (snd_cwnd) are in several places in the patch:

Right. I know what I've modified in that patch... ;-)

> 1. tcp_ipv4.c in function tcp_init_cwnd. since currently I am working
> with no metric this is irrelevant.

You managed to mix up this filename twice in a row in a different way :-). 
...That function resides in tcp_input.c instead. And changes in 
tcp_init_metrics are not completely irrelevant, even with no metrics it 
is executed as "no metrics" works a bit differenctly than you seem to 
think.

> 2. tcp_ipv4.c: in function tcp_v4_init_sock, which I applied
> 3. tcp_minisocks.c - which I applied
> . They patched kernel 2.6.18.8 which is very old and there were a lot
> of changes since that in the kernel tcp layer. I didn't test their
> patch but applying the same logic to 2.6.31.3 doesn't seems to change
> initial window size.
> 
> It is clear that in order to correctly make the change to the initial
> cwnd, one should apply changes to many more scenarios like setting the
> initial cwnd after idle time and more. I want to proceed step by step,
> first the naive changes to make it work.

Hmm, it now rings a bell... You cannot directly set initial window to a 
large value (or you can but that doesn't do anything on the wire) because 
you hit first the limit of receiver advertised window that applies 
auto-tuning (with a starting value was four IIRC).

-- 
 i.